This repository contains the official PyTorch implementation of the paper: Yichao Zhou, Shichen Liu, Yi Ma. "NeRD: Neural 3D Reflection Symmetry Detector". CVPR 2021.

Introduction

We present NeRD, a Neural 3D Reflection Symmetry Detector, which combines the strength of learning-based recognition and geometry-based reconstruction to accurately recover the normal direction of objects' mirror planes. NeRD uses coarse-to-fine strategy to enumerate the symmetry planes and then find the best ones by building 3D cost volumes to examine the intra-image pixel correspondence from the symmetry.

Qualitative Measures


ResNet Regression	NeRD	ResNet Regression	NeRD

Errors of the mirror plane are marked in red.

Code Structure

Below is a quick overview of the function of key files.

########################### Data ###########################
data/
    shapenet-r2n2/              # default folder for the shapenet dataset
    new_pix3d/                  # default folder for the pix3d dataset
logs/                           # default folder for storing the output during training
########################### Code ###########################
config/                         # neural network hyper-parameters and configurations
    shapenet.yaml               # example config for shapenet
    pix3d.yaml                  # example config for pix3d
misc/                           # misc scripts that are not important
    find-radius.py              # script for generating figure grids
sym/                            # sym module so you can "import sym" in other scripts
    models/                     # neural network architectures
        symmetry_net.py         # wrapper for loss
        mvsnet.py               # 3D hourglass
    config.py                   # global variables for configuration
    datasets.py                 # reading the training data
    trainer.py                  # general trainer
train.py                        # script for training the neural network
eval.py                         # script for evaluating a dataset from a checkpoint
plot-angle.py                   # script for ploting angle error curves
plot-depth.py                   # script for ploting depth error curves

Reproducing NeRD

Installation

For the ease of reproducibility, you are suggested to install miniconda before following executing the following commands.

git clone https://github.com/zhou13/nerd
cd nerd
conda create -y -n nerd
source activate nerd
conda install -y pyyaml docopt matplotlib scikit-image opencv tqdm
# Replace cudatoolkit=10.2 with your CUDA version: https://pytorch.org/get-started/
conda install -y pytorch cudatoolkit=10.2 -c pytorch
mkdir data logs results

Downloading the Processed Datasets

Make sure curl is installed on your system and execute

cd data
wget https://huggingface.co/yichaozhou/nerd/resolve/main/ShapeNet-R2N2.zip
wget https://huggingface.co/yichaozhou/nerd/resolve/main/new_pix3d.zip
unzip *.zip
rm *.zip
cd ..

If wget does not work for you, you can download the pre-processed datasets manually from Huggingface and proceed accordingly.

Training (Optional)

Execute the following commands to train the neural networks from scratch with four GPUs (specified by -d 0,1,2,3):

python ./train.py -d 0,1,2,3 --identifier baseline config/shapenet.yaml
python ./train.py -d 0,1,2,3 --identifier baseline config/pix3d.yaml

The checkpoints and logs will be written to logs/ accordingly.

Pre-Trained Models

cd logs/
wget https://huggingface.co/yichaozhou/nerd/resolve/main/Pre-Trained/201113-224159-ec0e932-pix3d.zip  # Pix3d/Symmetry
wget https://huggingface.co/yichaozhou/nerd/resolve/main/Pre-Trained/200610-234002-8ee0ad2-shapenet.zip  # ShapeNet/Depth
wget https://huggingface.co/yichaozhou/nerd/resolve/main/Pre-Trained/200513-030330-c8e671c-shapenet-finetune.zip  # ShapeNet/Symmetry
unzip *.zip
rm *.zip
cd ..

Alternatively, you can download our reference pre-trained from Huggingface.

Evaluation

To evaluate the models with coarse-to-fine inference for symmetry plane prediction and depth map estimation, execute

python eval.py -d 0 --output results/nerd.npz logs/<your-checkpoint>/config.yaml logs/<your-checkpoint>/checkpoint_latest.pth.tar

The error statistics are printed on the screen and the error metrics are stored in results/nerd.npz. To calculate the error metrics and plot the error-percentage curves, execute

python plot-angle.py
python plot-depth.py

Acknowledgement

This work is supported by the research grant from Sony, the ONR grant N00014-20-1-2002, and the joint Simons Foundation-NSF DMS grant 2031899. We also thank Li Yi from Google Research for his comments.

Citing NeRD

If you find NeRD useful in your research, please consider citing:

@inproceedings{zhou2021nerd,
    author = {Zhou, Yichao and Liu, Shichen and Ma, Yi},
    title = {{NeRD}: Neural 3D Reflection Symmetry Detector},
    year = {2021},
    booktitle = {CVPR},
}